When working with large codebases, you face two choices:

• The Manual Way: Copy-pasting files one by one (slow, loses context).
• The Standard RAG Way: Using chunk-based retrieval (fast, but often loses the "big picture" because it only sees tiny snippets of code).

✨ The Solution: Hierarchical RAG

This plugin implements a Hierarchical Retrieval-Augmented Generation (RAG) strategy. Instead of just searching for keywords in code chunks, it performs a two-tier analysis:

• Semantic Summarization: The plugin scans your local directory, strips out comments to save tokens, and uses an LLM to generate a high-level "architectural summary" of every file.
• Summary-Based Retrieval: When you ask a question, the plugin compares your query against the summaries (not the raw code). This allows it to identify the correct files based on their purpose.
• Full-Context Injection: Once the relevant files are identified, the plugin injects the entire content of those files into the prompt. This gives the LLM the full implementation details needed to provide accurate, working code.

🛠 Key Features

• 📂 Automatic Directory Scanning: Use the @dir:/path/to/folder trigger in chat to instantly bring a whole project into context.
• 🧹 Intelligent Token Optimization: Automatically strips comments (e.g., in .cs files) to maximize the amount of actual logic you can fit into the context window.
• 🧠 Hierarchical Search: Uses LLM-generated summaries to perform semantic searches, ensuring the LLM understands what a file does before it reads how it does it.
• ⚖️ Hybrid Decision Engine: Automatically decides whether to inject the full content (if the project is small) or switch to RAG (if the project is large) based on your specific injectionThreshold.
• 🎯 Configurable: Fine-tune your retrievalLimit, targetExtensions, and injectionThreshold via the LM Studio plugin settings.

🚀 Usage

• Install the plugin in LM Studio.
• Configure your target extensions (e.g., .cs, .ts, .py) in the settings.
• Chat with your code:

User: @dir:/Users/dev/MyProject Explain how the authentication flow works in this app.

The Plugin will:

• Scan MyProject.
• Summarize the files.
• Realize AuthService.cs and LoginController.cs are the most relevant.
• Inject the full content of those files into the prompt.
• Answer your question with full awareness of your code.

🚧 Roadmap (WIP)

This project is currently under active development. My current focus is:

• Incremental Summarization (Current Focus): Implementing a caching layer so that only modified files are re-summarized, making the @dir: command nearly instantaneous.
• Graph-Based Relationships: Moving from simple summaries to a true GraphRAG that maps function calls and class inheritance.
• Multi-Language Support: Expanding the comment-stripping logic to support Python, JavaScript, and C++.

🛠 Tech Stack

• Runtime: Node.js / TypeScript
• SDK: @lmstudio/sdk
• Embeddings: Nomic Embed Text (via LM Studio)

📄 License

This project is licensed under the MIT License.

Built with ❤️ for developers who want to talk to their code.